图文精华

About云-梭伦科技»专题 › 技术学习(版主发帖区) › 大数据学习 › 数据湖 › 技术应用› 使用 Flink Hudi 构建流式数据湖
分享

使用 Flink Hudi 构建流式数据湖

Mirinda 2021-9-2 14:17:50 发表于 技术应用 [显示全部楼层] 只看大图 回帖奖励 阅读模式 关闭右栏 0 3721
本帖最后由 Mirinda 于 2021-9-2 14:47 编辑

问题导读:
1.现实中有哪些案例属于近实时的范畴?
2.增量模型需要解决那些核心问题?
3.一张hudi数据湖表具有那些形态?


一、背景近实时
从 2016 年开始，Apache Hudi 社区就开始通过 Hudi 的 UPSERT 能力探索近实时场景的使用案例 [1]。通过 MR/Spark 的批处理模型，用户可以实现小时级别的数据注入 HDFS/OSS。在纯实时场景，用户通过流计算引擎 Flink + KV/OLAP 存储的架构可以实现端到端的秒级 (5分钟级) 实时分析。然而在秒级 (5分钟级) 到小时级时的场景还存在大量的用例，我们称之为 NEAR-REAL-TIME (近实时)。
a1.png
在实践中有大量的案例都属于近实时的范畴：
  • 分钟级别的大屏；
  • 各种 BI 分析 (OLAP)；
  • 机器学习分钟级别的特征提取。

增量计算
解决近实时的方案当前是比较开放的。
  • 流处理的时延低，但是 SQL 的 pattern 比较固定，查询端的能力（索引、ad hoc）欠缺；
  • 批处理的数仓能力丰富但是数据时延大。

于是 Hudi 社区提出基于 mini-batch 的增量计算模型：
增量数据集 => 增量计算结果 merge 已存结果 => 外存
这套模型通过湖存储的 snapshot 拉取增量的数据集 (两个 commits 之前的数据集)，通过 Spark/Hive 等批处理框架计算增量的结果 (比如简单的 count) 再 merge 到已存结果中。

核心问题
增量模型需要解决的核心问题：
  • UPSERT 能力：类似 KUDU 和 Hive ACID，Hudi 也提供了分钟级的更新能力；
  • 增量消费：Hudi 通过湖存储的多 snapshots 提供增量拉取。
基于 mini-batch 的增量计算模型可以提升部分场景的时延、节省计算成本，但有一个很大的限制：对 SQL 的 pattern 有要求。因为计算走的是批，批计算本身不维护状态，这就要求计算的指标能够比较方便地 merge，简单的 count、sum 可以做，但是 avg、count distinct 这些还是需要拉取全量数据重算。
随着流计算和实时数仓的普及，Hudi 社区也在积极的拥抱变化，通过流计算对原有基于 mini-batch 的增量计算模型不断优化演进：在 0.7 版本引入了流式数据入湖，在 0.9 版本支持了原生的 CDC format。

二、增量 ETLDB 数据入湖
随着 CDC 技术的成熟，debezium 这样的 CDC 工具越来越流行，Hudi 社区也先后集成了流写，流读的能力。用户可以通过 Flink SQL 将 CDC 数据实时写入 Hudi 存储：
a2.png
  • 用户既可以通过 Flink CDC connector 直接将 DB 数据导入 Hudi；
  • 也可以先将 CDC 数据导入 Kafka，再通过 Kafka connector 导入 Hudi。

第二种方案的容错和扩展性会好一些。
数据湖 CDC
在即将发布的 0.9 版本，Hudi 原生支持 CDC format，一条 record 的所有变更记录都可以保存，基于此，Hudi 和流计算系统结合的更加完善，可以流式读取 CDC 数据 [2]：
a3.png
源头 CDC 流的所有消息变更都在入湖之后保存下来，被用于流式消费。Flink 的有状态计算实时累加计算结果 (state)，通过流式写 Hudi 将计算的变更同步到 Hudi 湖存储，之后继续对接 Flink 流式消费 Hudi 存储的 changelog， 实现下一层级的有状态计算。近实时端到端 ETL pipeline：
a4.png
这套架构将端到端的 ETL 时延缩短到分钟级，并且每一层的存储格式都可以通过 compaction 压缩成列存（Parquet、ORC）以提供 OLAP 分析能力，由于数据湖的开放性，压缩后的格式可以对接各种查询引擎：Flink、Spark、Presto、Hive 等。

一张 Hudi 数据湖表具备两种形态：
  • 表形态：查询最新的快照结果，同时提供高效的列存格式
  • 流形态：流式消费变更，可以指定任意点位流读之后的 changelog

三、演示
我们通过一段 Demo 演示 Hudi 表的两种形态。
环境准备
  • Flink SQL Client
  • Hudi master 打包 hudi-flink-bundle jar
  • Flink 1.13.1

这里提前准备一段 debezium-json 格式的 CDC 数据
  1. {"before":null,"after":{"id":101,"ts":1000,"name":"scooter","description":"Small 2-wheel scooter","weight":3.140000104904175},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606100,"transaction":null}

  2. {"before":null,"after":{"id":102,"ts":2000,"name":"car battery","description":"12V car battery","weight":8.100000381469727},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  3. {"before":null,"after":{"id":103,"ts":3000,"name":"12-pack drill bits","description":"12-pack of drill bits with sizes ranging from #40 to #3","weight":0.800000011920929},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  4. {"before":null,"after":{"id":104,"ts":4000,"name":"hammer","description":"12oz carpenter's hammer","weight":0.75},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  5. {"before":null,"after":{"id":105,"ts":5000,"name":"hammer","description":"14oz carpenter's hammer","weight":0.875},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  6. {"before":null,"after":{"id":106,"ts":6000,"name":"hammer","description":"16oz carpenter's hammer","weight":1},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  7. {"before":null,"after":{"id":107,"ts":7000,"name":"rocks","description":"box of assorted rocks","weight":5.300000190734863},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  8. {"before":null,"after":{"id":108,"ts":8000,"name":"jacket","description":"water resistent black wind breaker","weight":0.10000000149011612},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  9. {"before":null,"after":{"id":109,"ts":9000,"name":"spare tire","description":"24 inch spare tire","weight":22.200000762939453},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":0,"snapshot":"true","db":"inventory","table":"products","server_id":0,"gtid":null,"file":"mysql-bin.000003","pos":154,"row":0,"thread":null,"query":null},"op":"c","ts_ms":1589355606101,"transaction":null}

  10. {"before":{"id":106,"ts":6000,"name":"hammer","description":"16oz carpenter's hammer","weight":1},"after":{"id":106,"ts":10000,"name":"hammer","description":"18oz carpenter hammer","weight":1},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":1589361987000,"snapshot":"false","db":"inventory","table":"products","server_id":223344,"gtid":null,"file":"mysql-bin.000003","pos":362,"row":0,"thread":2,"query":null},"op":"u","ts_ms":1589361987936,"transaction":null}

  11. {"before":{"id":107,"ts":7000,"name":"rocks","description":"box of assorted rocks","weight":5.300000190734863},"after":{"id":107,"ts":11000,"name":"rocks","description":"box of assorted rocks","weight":5.099999904632568},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":1589362099000,"snapshot":"false","db":"inventory","table":"products","server_id":223344,"gtid":null,"file":"mysql-bin.000003","pos":717,"row":0,"thread":2,"query":null},"op":"u","ts_ms":1589362099505,"transaction":null}

  12. {"before":null,"after":{"id":110,"ts":12000,"name":"jacket","description":"water resistent white wind breaker","weight":0.20000000298023224},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":1589362210000,"snapshot":"false","db":"inventory","table":"products","server_id":223344,"gtid":null,"file":"mysql-bin.000003","pos":1068,"row":0,"thread":2,"query":null},"op":"c","ts_ms":1589362210230,"transaction":null}

  13. {"before":null,"after":{"id":111,"ts":13000,"name":"scooter","description":"Big 2-wheel scooter ","weight":5.179999828338623},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":1589362243000,"snapshot":"false","db":"inventory","table":"products","server_id":223344,"gtid":null,"file":"mysql-bin.000003","pos":1394,"row":0,"thread":2,"query":null},"op":"c","ts_ms":1589362243428,"transaction":null}

  14. {"before":{"id":110,"ts":12000,"name":"jacket","description":"water resistent white wind breaker","weight":0.20000000298023224},"after":{"id":110,"ts":14000,"name":"jacket","description":"new water resistent white wind breaker","weight":0.5},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":1589362293000,"snapshot":"false","db":"inventory","table":"products","server_id":223344,"gtid":null,"file":"mysql-bin.000003","pos":1707,"row":0,"thread":2,"query":null},"op":"u","ts_ms":1589362293539,"transaction":null}

  15. {"before":{"id":111,"ts":13000,"name":"scooter","description":"Big 2-wheel scooter ","weight":5.179999828338623},"after":{"id":111,"ts":15000,"name":"scooter","description":"Big 2-wheel scooter ","weight":5.170000076293945},"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":1589362330000,"snapshot":"false","db":"inventory","table":"products","server_id":223344,"gtid":null,"file":"mysql-bin.000003","pos":2090,"row":0,"thread":2,"query":null},"op":"u","ts_ms":1589362330904,"transaction":null}

  16. {"before":{"id":111,"ts":16000,"name":"scooter","description":"Big 2-wheel scooter ","weight":5.170000076293945},"after":null,"source":{"version":"1.1.1.Final","connector":"mysql","name":"dbserver1","ts_ms":1589362344000,"snapshot":"false","db":"inventory","table":"products","server_id":223344,"gtid":null,"file":"mysql-bin.000003","pos":2443,"row":0,"thread":2,"query":null},"op":"d","ts_ms":1589362344455,"transaction":null}
复制代码

通过 Flink SQL Client 创建表用来读取 CDC 数据文件
  1. Flink SQL> CREATE TABLE debezium_source(
复制代码
执行 SELECT 观察结果，可以看到一共有 20 条记录，中间有一些 UPDATE s，最后一条消息是 DELETE
Flink SQL> select * from debezium_source;
a5.png

创建 Hudi 表，这里设置表的形态为 MERGE_ON_READ 并且打开 changelog 模式属性 changelog.enabled
  1. Flink SQL> CREATE TABLE hoodie_table(
复制代码

查询
通过 INSERT 语句将数据导入 Hudi，开启流读模式，并执行查询观察结果

a6.png


可以看到 Hudi 保留了每行的变更记录，包括 change log 的 operation 类型，这里我们打开 TABLE HINTS 功能，方便动态设置表参数。
继续使用 batch 读模式，执行查询观察输出结果，可以看到中间的变更被合并。

a7.png


聚合
Bounded Source 读模式下计算 count(*)

a8.png
a9.png


Streaming 读模式下计算 count(*)

a10.png
a11.png


可以看到 batch 和 streaming 模式下的计算结果是一致的。
当前的数据湖 CDC format 还处在快速迭代期，社区也在积极推动生产场景，对 Hudi 场景和案例感兴趣的同学可以扫码加群。

最新经典文章，欢迎关注公众号









回复

使用道具 举报

自定义广告语

没找到任何评论，期待你打破沉寂

高级模式
B Color Image Link Quote Code Smilies
您需要登录后才可以回帖 登录 | 立即注册

本版积分规则

发表新帖
Mirinda

超级版主

关注

  • 123

    主题

  • 125

    帖子

  • 11

    粉丝

TA的主题
关闭

推荐上一条 /2 下一条

关于我们· 联系我们· 加入我们· 小黑屋· 合作伙伴   

Copyright © 2001-2024 About云-梭伦科技 Powered by Discuz! X3.4 Licensed Discuz Team.

简书 / Smrz 京ICP备2020039040号 Wxb 简书网举报电话：021-34700000